]>
git.ipfire.org Git - thirdparty/mdadm.git/blob - managemon.c
2 * mdmon - monitor external metadata arrays
4 * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de>
5 * Copyright (C) 2007-2009 Intel Corporation
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
16 * You should have received a copy of the GNU General Public License along with
17 * this program; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
22 * The management thread for monitoring active md arrays.
23 * This thread does things which might block such as memory
27 * - Find out about new arrays in this container.
28 * Allocate the data structures and open the files.
30 * For this we watch /proc/mdstat and find new arrays with
31 * metadata type that confirms sharing. e.g. "md4"
32 * When we find a new array we slip it into the list of
33 * arrays and signal 'monitor' by writing to a pipe.
35 * - Respond to reshape requests by allocating new data structures
36 * and opening new files.
38 * These come as a change to raid_disks. We allocate a new
39 * version of the data structures and slip it into the list.
40 * 'monitor' will notice and release the old version.
41 * Changes to level, chunksize, layout.. do not need re-allocation.
42 * Reductions in raid_disks don't really either, but we handle
43 * them the same way for consistency.
45 * - When a device is added to the container, we add it to the metadata
48 * - Deal with degraded array
49 * We only do this when first noticing the array is degraded.
50 * This can be when we first see the array, when sync completes or
51 * when recovery completes.
53 * Check if number of failed devices suggests recovery is needed, and
55 * Ask metadata to allocate a spare device
56 * Add device as not in_sync and give a role
58 * Open sysfs files and pass to monitor.
59 * Make sure that monitor Starts recovery....
61 * - Pass on metadata updates from external programs such as
62 * mdadm creating a new array.
65 * It might involve adding a new array or changing the status of
66 * a spare, or any reconfig that the kernel doesn't get involved in.
68 * The required updates are received via a named pipe. There will
69 * be one named pipe for each container. Each message contains a
70 * sync marker: 0x5a5aa5a5, A byte count, and the message. This is
71 * passed to the metadata handler which will interpret and process it.
72 * For 'DDF' messages are internal data blocks with the leading
73 * 'magic number' signifying what sort of data it is.
78 * We select on /proc/mdstat and the named pipe.
79 * We create new arrays or updated version of arrays and slip
80 * them into the head of the list, then signal 'monitor' via a pipe write.
81 * 'monitor' will notice and place the old array on a return list.
82 * Metadata updates are placed on a queue just like they arrive
83 * from the named pipe.
85 * When new arrays are found based on correct metadata string, we
86 * need to identify them with an entry in the metadata. Maybe we require
87 * the metadata to be mdX/NN when NN is the index into an appropriate table.
93 * - Watch for spares to be added to the container, and write updated
95 * - Watch for new arrays using this container, confirm they match metadata
96 * and if so, start monitoring them
97 * - Watch for spares being added to monitored arrays. This shouldn't
98 * happen, as we should do all the adding. Just remove them.
99 * - Watch for change in raid-disks, chunk-size, etc. Update metadata and
107 #include <sys/syscall.h>
108 #include <sys/socket.h>
111 static void close_aa(struct active_array
*aa
)
115 for (d
= aa
->info
.devs
; d
; d
= d
->next
) {
116 close(d
->recovery_fd
);
120 close(aa
->action_fd
);
121 close(aa
->info
.state_fd
);
122 close(aa
->resync_start_fd
);
123 close(aa
->metadata_fd
);
124 close(aa
->sync_completed_fd
);
127 static void free_aa(struct active_array
*aa
)
129 /* Note that this doesn't close fds if they are being used
130 * by a clone. ->container will be set for a clone
132 dprintf("%s: devnum: %d\n", __func__
, aa
->devnum
);
135 while (aa
->info
.devs
) {
136 struct mdinfo
*d
= aa
->info
.devs
;
137 aa
->info
.devs
= d
->next
;
143 static struct active_array
*duplicate_aa(struct active_array
*aa
)
145 struct active_array
*newa
= malloc(sizeof(*newa
));
146 struct mdinfo
**dp1
, **dp2
;
150 newa
->replaces
= NULL
;
151 newa
->info
.next
= NULL
;
153 dp2
= &newa
->info
.devs
;
155 for (dp1
= &aa
->info
.devs
; *dp1
; dp1
= &(*dp1
)->next
) {
157 if ((*dp1
)->state_fd
< 0)
160 d
= malloc(sizeof(*d
));
170 static void wakeup_monitor(void)
172 /* tgkill(getpid(), mon_tid, SIGUSR1); */
174 syscall(SYS_tgkill
, pid
, mon_tid
, SIGUSR1
);
177 static void remove_old(void)
180 discard_this
->next
= NULL
;
181 free_aa(discard_this
);
182 if (pending_discard
== discard_this
)
183 pending_discard
= NULL
;
189 static void replace_array(struct supertype
*container
,
190 struct active_array
*old
,
191 struct active_array
*new)
193 /* To replace an array, we add it to the top of the list
194 * marked with ->replaces to point to the original.
195 * 'monitor' will take the original out of the list
196 * and put it on 'discard_this'. We take it from there
200 while (pending_discard
) {
201 while (discard_this
== NULL
)
205 pending_discard
= old
;
207 new->next
= container
->arrays
;
208 container
->arrays
= new;
212 struct metadata_update
*update_queue
= NULL
;
213 struct metadata_update
*update_queue_handled
= NULL
;
214 struct metadata_update
*update_queue_pending
= NULL
;
216 static void free_updates(struct metadata_update
**update
)
219 struct metadata_update
*this = *update
;
220 void **space_list
= this->space_list
;
222 *update
= this->next
;
226 void *space
= space_list
;
227 space_list
= *space_list
;
234 void check_update_queue(struct supertype
*container
)
236 free_updates(&update_queue_handled
);
238 if (update_queue
== NULL
&&
239 update_queue_pending
) {
240 update_queue
= update_queue_pending
;
241 update_queue_pending
= NULL
;
246 static void queue_metadata_update(struct metadata_update
*mu
)
248 struct metadata_update
**qp
;
250 qp
= &update_queue_pending
;
252 qp
= & ((*qp
)->next
);
256 static void add_disk_to_container(struct supertype
*st
, struct mdinfo
*sd
)
260 struct supertype
*st2
;
261 struct metadata_update
*update
= NULL
;
263 mdu_disk_info_t dk
= {
265 .major
= sd
->disk
.major
,
266 .minor
= sd
->disk
.minor
,
271 dprintf("%s: add %d:%d to container\n",
272 __func__
, sd
->disk
.major
, sd
->disk
.minor
);
277 sprintf(nm
, "%d:%d", sd
->disk
.major
, sd
->disk
.minor
);
278 dfd
= dev_open(nm
, O_RDWR
);
282 /* Check the metadata and see if it is already part of this
286 if (st2
->ss
->load_super(st2
, dfd
, NULL
) == 0) {
287 st2
->ss
->getinfo_super(st
, &info
, NULL
);
288 if (st
->ss
->compare_super(st
, st2
) == 0 &&
289 info
.disk
.raid_disk
>= 0) {
290 /* Looks like a good member of array.
292 * mdadm will incorporate any parts into
295 st2
->ss
->free_super(st2
);
299 st2
->ss
->free_super(st2
);
301 st
->update_tail
= &update
;
302 st
->ss
->add_to_super(st
, &dk
, dfd
, NULL
);
303 st
->ss
->write_init_super(st
);
304 queue_metadata_update(update
);
305 st
->update_tail
= NULL
;
309 * Create and queue update structure about the removed disks.
310 * The update is prepared by super type handler and passed to the monitor
313 static void remove_disk_from_container(struct supertype
*st
, struct mdinfo
*sd
)
315 struct metadata_update
*update
= NULL
;
316 mdu_disk_info_t dk
= {
318 .major
= sd
->disk
.major
,
319 .minor
= sd
->disk
.minor
,
323 dprintf("%s: remove %d:%d from container\n",
324 __func__
, sd
->disk
.major
, sd
->disk
.minor
);
326 st
->update_tail
= &update
;
327 st
->ss
->remove_from_super(st
, &dk
);
328 /* FIXME this write_init_super shouldn't be here.
329 * We have it after add_to_super to write to new device,
330 * but with 'remove' we don't ant to write to that device!
332 st
->ss
->write_init_super(st
);
333 queue_metadata_update(update
);
334 st
->update_tail
= NULL
;
337 static void manage_container(struct mdstat_ent
*mdstat
,
338 struct supertype
*container
)
340 /* Of interest here are:
341 * - if a new device has been added to the container, we
342 * add it to the array ignoring any metadata on it.
343 * - if a device has been removed from the container, we
344 * remove it from the device list and update the metadata.
345 * FIXME should we look for compatible metadata and take hints
346 * about spare assignment.... probably not.
348 if (mdstat
->devcnt
!= container
->devcnt
) {
349 struct mdinfo
**cdp
, *cd
, *di
, *mdi
;
352 /* read /sys/block/NAME/md/dev-??/block/dev to find out
353 * what is there, and compare with container->info.devs
354 * To see what is removed and what is added.
355 * These need to be remove from, or added to, the array
357 mdi
= sysfs_read(-1, mdstat
->devnum
, GET_DEVS
);
359 /* invalidate the current count so we can try again */
360 container
->devcnt
= -1;
364 /* check for removals */
365 for (cdp
= &container
->devs
; *cdp
; ) {
367 for (di
= mdi
->devs
; di
; di
= di
->next
)
368 if (di
->disk
.major
== (*cdp
)->disk
.major
&&
369 di
->disk
.minor
== (*cdp
)->disk
.minor
) {
376 remove_disk_from_container(container
, cd
);
382 /* check for additions */
383 for (di
= mdi
->devs
; di
; di
= di
->next
) {
384 for (cd
= container
->devs
; cd
; cd
= cd
->next
)
385 if (di
->disk
.major
== cd
->disk
.major
&&
386 di
->disk
.minor
== cd
->disk
.minor
)
389 struct mdinfo
*newd
= malloc(sizeof(*newd
));
392 container
->devcnt
= -1;
396 add_disk_to_container(container
, newd
);
400 container
->devcnt
= mdstat
->devcnt
;
404 static int disk_init_and_add(struct mdinfo
*disk
, struct mdinfo
*clone
,
405 struct active_array
*aa
)
411 disk
->recovery_fd
= sysfs_open(aa
->devnum
, disk
->sys_name
, "recovery_start");
412 disk
->state_fd
= sysfs_open(aa
->devnum
, disk
->sys_name
, "state");
413 disk
->prev_state
= read_dev_state(disk
->state_fd
);
414 disk
->curr_state
= disk
->prev_state
;
415 disk
->next
= aa
->info
.devs
;
416 aa
->info
.devs
= disk
;
421 static void manage_member(struct mdstat_ent
*mdstat
,
422 struct active_array
*a
)
424 /* Compare mdstat info with known state of member array.
425 * We do not need to look for device state changes here, that
426 * is dealt with by the monitor.
428 * If a reshape is being requested, monitor will have noticed
429 * that sync_action changed and will have set check_reshape.
430 * We just need to see if new devices have appeared. All metadata
431 * updates will already have been processed.
433 * We also want to handle degraded arrays here by
434 * trying to find and assign a spare.
435 * We do that whenever the monitor tells us too.
439 struct supertype
*container
= a
->container
;
441 if (container
== NULL
)
442 /* Raced with something */
446 a
->info
.array
.raid_disks
= mdstat
->raid_disks
;
450 if (sysfs_get_str(&a
->info
, NULL
, "metadata_version", buf
, sizeof(buf
)) > 0)
451 frozen
= buf
[9] == '-';
453 frozen
= 1; /* can't read metadata_version assume the worst */
455 /* If sync_action is not 'idle' then don't try recovery now */
457 && sysfs_get_str(&a
->info
, NULL
, "sync_action", buf
, sizeof(buf
)) > 0
458 && strncmp(buf
, "idle", 4) != 0)
462 int level
= map_name(pers
, mdstat
->level
);
463 if (level
== 0 || level
== LEVEL_LINEAR
) {
468 else if (a
->info
.array
.level
!= level
&& level
> 0) {
469 struct active_array
*newa
= duplicate_aa(a
);
471 newa
->info
.array
.level
= level
;
472 replace_array(container
, a
, newa
);
478 /* We don't check the array while any update is pending, as it
479 * might container a change (such as a spare assignment) which
480 * could affect our decisions.
482 if (a
->check_degraded
&& !frozen
&&
483 update_queue
== NULL
&& update_queue_pending
== NULL
) {
484 struct metadata_update
*updates
= NULL
;
485 struct mdinfo
*newdev
= NULL
;
486 struct active_array
*newa
;
489 a
->check_degraded
= 0;
491 /* The array may not be degraded, this is just a good time
494 newdev
= container
->ss
->activate_spare(a
, &updates
);
498 newa
= duplicate_aa(a
);
501 /* prevent the kernel from activating the disk(s) before we
504 sysfs_set_str(&a
->info
, NULL
, "sync_action", "frozen");
506 /* Add device to array and set offset/size/slot.
507 * and open files for each newdev */
508 for (d
= newdev
; d
; d
= d
->next
) {
511 newd
= malloc(sizeof(*newd
));
514 if (sysfs_add_disk(&newa
->info
, d
, 0) < 0) {
518 disk_init_and_add(newd
, d
, newa
);
520 queue_metadata_update(updates
);
522 replace_array(container
, a
, newa
);
523 sysfs_set_str(&a
->info
, NULL
, "sync_action", "recover");
530 free_updates(&updates
);
533 if (a
->check_reshape
) {
534 /* mdadm might have added some devices to the array.
535 * We want to disk_init_and_add any such device to a
536 * duplicate_aa and replace a with that.
537 * mdstat doesn't have enough info so we sysfs_read
538 * and look for new stuff.
540 struct mdinfo
*info
, *d
, *d2
, *newd
;
541 unsigned long long array_size
;
542 struct active_array
*newa
= NULL
;
543 a
->check_reshape
= 0;
544 info
= sysfs_read(-1, mdstat
->devnum
,
545 GET_DEVS
|GET_OFFSET
|GET_SIZE
|GET_STATE
);
548 for (d
= info
->devs
; d
; d
= d
->next
) {
549 if (d
->disk
.raid_disk
< 0)
551 for (d2
= a
->info
.devs
; d2
; d2
= d2
->next
)
552 if (d2
->disk
.raid_disk
==
556 /* already have this one */
559 newa
= duplicate_aa(a
);
563 newd
= malloc(sizeof(*newd
));
566 disk_init_and_add(newd
, d
, newa
);
568 if (sysfs_get_ll(info
, NULL
, "array_size", &array_size
) == 0
569 && a
->info
.custom_array_size
> array_size
*2) {
570 sysfs_set_num(info
, NULL
, "array_size",
571 a
->info
.custom_array_size
/2);
576 replace_array(container
, a
, newa
);
580 static int aa_ready(struct active_array
*aa
)
583 int level
= aa
->info
.array
.level
;
585 for (d
= aa
->info
.devs
; d
; d
= d
->next
)
589 if (aa
->info
.state_fd
< 0)
592 if (level
> 0 && (aa
->action_fd
< 0 || aa
->resync_start_fd
< 0))
601 static void manage_new(struct mdstat_ent
*mdstat
,
602 struct supertype
*container
,
603 struct active_array
*victim
)
605 /* A new array has appeared in this container.
606 * Hopefully it is already recorded in the metadata.
607 * Check, then create the new array to report it to
611 struct active_array
*new;
612 struct mdinfo
*mdi
, *di
;
618 /* check if array is ready to be monitored */
619 if (!mdstat
->active
|| !mdstat
->level
)
621 if (strcmp(mdstat
->level
, "raid0") == 0 ||
622 strcmp(mdstat
->level
, "linear") == 0)
625 mdi
= sysfs_read(-1, mdstat
->devnum
,
626 GET_LEVEL
|GET_CHUNK
|GET_DISKS
|GET_COMPONENT
|
627 GET_DEGRADED
|GET_DEVS
|GET_OFFSET
|GET_SIZE
|GET_STATE
);
629 new = malloc(sizeof(*new));
638 memset(new, 0, sizeof(*new));
640 new->devnum
= mdstat
->devnum
;
641 strcpy(new->info
.sys_name
, devnum2devname(new->devnum
));
643 new->prev_state
= new->curr_state
= new->next_state
= inactive
;
644 new->prev_action
= new->curr_action
= new->next_action
= idle
;
646 new->container
= container
;
648 inst
= to_subarray(mdstat
, container
->devname
);
650 new->info
.array
= mdi
->array
;
651 new->info
.component_size
= mdi
->component_size
;
653 for (i
= 0; i
< new->info
.array
.raid_disks
; i
++) {
654 struct mdinfo
*newd
= malloc(sizeof(*newd
));
656 for (di
= mdi
->devs
; di
; di
= di
->next
)
657 if (i
== di
->disk
.raid_disk
)
660 if (disk_init_and_add(newd
, di
, new) != 0) {
665 if (failed
> new->info
.array
.failed_disks
) {
666 /* we cannot properly monitor without all working disks */
667 new->container
= NULL
;
673 new->action_fd
= sysfs_open(new->devnum
, NULL
, "sync_action");
674 new->info
.state_fd
= sysfs_open(new->devnum
, NULL
, "array_state");
675 new->resync_start_fd
= sysfs_open(new->devnum
, NULL
, "resync_start");
676 new->metadata_fd
= sysfs_open(new->devnum
, NULL
, "metadata_version");
677 new->sync_completed_fd
= sysfs_open(new->devnum
, NULL
, "sync_completed");
678 dprintf("%s: inst: %d action: %d state: %d\n", __func__
, atoi(inst
),
679 new->action_fd
, new->info
.state_fd
);
681 /* reshape_position is set by mdadm in sysfs
682 * read this information for new arrays only (empty victim)
684 if ((victim
== NULL
) &&
685 (sysfs_get_str(mdi
, NULL
, "sync_action", buf
, 40) > 0) &&
686 (strncmp(buf
, "reshape", 7) == 0)) {
687 if (sysfs_get_ll(mdi
, NULL
, "reshape_position",
688 &new->last_checkpoint
) != 0)
689 new->last_checkpoint
= 0;
691 int data_disks
= mdi
->array
.raid_disks
;
692 if (mdi
->array
.level
== 4 || mdi
->array
.level
== 5)
694 if (mdi
->array
.level
== 6)
697 new->last_checkpoint
/= data_disks
;
699 dprintf("mdmon: New monitored array is under reshape.\n"
700 " Last checkpoint is: %llu\n",
701 new->last_checkpoint
);
706 /* if everything checks out tell the metadata handler we want to
707 * manage this instance
709 if (!aa_ready(new) || container
->ss
->open_new(container
, new, inst
) < 0) {
710 fprintf(stderr
, "mdmon: failed to monitor %s\n",
711 mdstat
->metadata_version
);
712 new->container
= NULL
;
715 replace_array(container
, victim
, new);
717 new->check_degraded
= 1;
718 manage_member(mdstat
, new);
723 void manage(struct mdstat_ent
*mdstat
, struct supertype
*container
)
725 /* We have just read mdstat and need to compare it with
726 * the known active arrays.
727 * Arrays with the wrong metadata are ignored.
730 for ( ; mdstat
; mdstat
= mdstat
->next
) {
731 struct active_array
*a
;
732 if (mdstat
->devnum
== container
->devnum
) {
733 manage_container(mdstat
, container
);
736 if (!is_container_member(mdstat
, container
->devname
))
737 /* Not for this array */
739 /* Looks like a member of this container */
740 for (a
= container
->arrays
; a
; a
= a
->next
) {
741 if (mdstat
->devnum
== a
->devnum
) {
742 if (a
->container
&& a
->to_remove
== 0)
743 manage_member(mdstat
, a
);
747 if (a
== NULL
|| !a
->container
)
748 manage_new(mdstat
, container
, a
);
752 static void handle_message(struct supertype
*container
, struct metadata_update
*msg
)
754 /* queue this metadata update through to the monitor */
756 struct metadata_update
*mu
;
759 while (update_queue_pending
|| update_queue
) {
760 check_update_queue(container
);
764 if (msg
->len
== 0) { /* ping_monitor */
767 cnt
= monitor_loop_cnt
;
769 cnt
+= 2; /* wait until next pselect */
771 cnt
+= 3; /* wait for 2 pselects */
774 while (monitor_loop_cnt
- cnt
< 0)
776 } else if (msg
->len
== -1) { /* ping_manager */
777 struct mdstat_ent
*mdstat
= mdstat_read(1, 0);
779 manage(mdstat
, container
);
781 } else if (!sigterm
) {
782 mu
= malloc(sizeof(*mu
));
787 mu
->space_list
= NULL
;
789 if (container
->ss
->prepare_update
)
790 container
->ss
->prepare_update(container
, mu
);
791 queue_metadata_update(mu
);
795 void read_sock(struct supertype
*container
)
798 struct metadata_update msg
;
801 int tmo
= 3; /* 3 second timeout before hanging up the socket */
803 fd
= accept(container
->sock
, NULL
, NULL
);
807 fl
= fcntl(fd
, F_GETFL
, 0);
809 fcntl(fd
, F_SETFL
, fl
);
814 /* read and validate the message */
815 if (receive_message(fd
, &msg
, tmo
) == 0) {
816 handle_message(container
, &msg
);
818 /* ping reply with version */
820 msg
.len
= strlen(Version
) + 1;
821 if (send_message(fd
, &msg
, tmo
) < 0)
823 } else if (ack(fd
, tmo
) < 0)
828 } while (!terminate
);
834 int manager_ready
= 0;
835 void do_manager(struct supertype
*container
)
837 struct mdstat_ent
*mdstat
;
840 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
841 sigdelset(&set
, SIGUSR1
);
842 sigdelset(&set
, SIGTERM
);
849 /* Can only 'manage' things if 'monitor' is not making
850 * structural changes to metadata, so need to check
853 if (update_queue
== NULL
) {
854 mdstat
= mdstat_read(1, 0);
856 manage(mdstat
, container
);
858 read_sock(container
);
864 check_update_queue(container
);
871 if (update_queue
== NULL
)
872 mdstat_wait_fd(container
->sock
, &set
);
874 /* If an update is happening, just wait for signal */
875 pselect(0, NULL
, NULL
, NULL
, NULL
, &set
);